Thanks for the overview of what issues you
faced and how you handled them.
The term you used ( “observable
quality space” ) as partitioned by said quality space caught my
attention. Are you viewing implementing that as a fixed set of columns (properties,
qualities) which are key columns in the rough set implementation?
Mapping the rough sets idea into SQL seems
like a very direct mapping if that is true:
1. Make every
element of the set of observable qualities represented by a unique key column;
2. Have the rough
set represented in a single table with key columns (observable properties) and
nonkey columns (extra contextual unobservable properties);
3. Define two
views – the upper and lower crisp sets – for exploring the table
4. Define a risk
view – the rows that fall into the upper boundary but not in the lower
boundary – that can be explored as the set of outlier cases.
Am I interpreting your results correctly
or is there more complexity to it than I anticipate?
Rich AT EnglishLogicKernel DOT com
9 4 9 \ 5 2 5 - 5 7 1 2
[mailto:ontolog-forum-bounces@xxxxxxxxxxxxxxxx] On Behalf Of Tara Athan
Sent: Friday, January 14, 2011
To: Rich Cooper
Cc: '[ontolog-forum] '
Subject: Re: [ontolog-forum]
Ontology of Rough Sets
Rich Cooper wrote:
I don’t recall any specific
reference on rough sets – I think I first read something in the IEEE
suite of pubs, maybe ten years ago.
Rough sets should be fairly easy to
implement in SQL, so it would be interesting to know what components you were
considering, and whether they met your needs.
Well, my needs were academic, I was using it as a
learning example. So I built a concrete boat, and I am reluctant to try to
float it in public. In any case, I did not create classes named after the
terms in rough set mathematics and try to formalize their definitions - I don't
see any point in that. What I did do was play around with ontology design
patterns inspired by rough set mathematics and multi-valued logic.
The design pattern is that rough classes are characterized, vaguely, by a pair
of crisp classes. A crisp class has a definition whose restrictions arise from
a finite partition of an observable quality space, which is a fairly typical
ontology design pattern to begin with. The pair of crisp classes are defined
using the same partitions, and the one corresponding to the lower approximation
must be a subclass of the upper approximation. The rough class is a super-class
of the lower approximation and a sub-class of the upper approximation. In
an application, it would be precisely defined in terms of some target quality
space which is not directly observable (in the setting where the ontology is to
be applied, perhaps clinical diagnosis, or sensors for fault-detection). The
lower/upper approximation give a vague description of this class in terms of
observable qualities. These approximations would be determined by statistics on
a sample where the target quality is known, say from a research study, or by
I don't think there is anything particularly profound in this - for all I know
people are already doing this and calling it something else. The connection to
rough set theory just gives us a vocabulary to talk about it.
If any reference info comes back to me,
I’ll post it.